NazraJN/Kenya-Monthly-Temperature-Time-series-forecasting

A time series forecasting project comparing SARIMA and Holt-Winters models to analyze temperature trends in Kenya using real-world climate data. # Kenya Temperature Forecasting - SARIMA & Holt-Winters ## Overview This project presents a time series analysis of Kenya's monthly average temperature from 1991 to 2016. The objective is to model and forecast temperature for a 12-month horizon using classical statistical methods. Two models are implemented and compared: - Seasonal ARIMA (SARIMA) - Holt-Winters Exponential Smoothing --- ## Dataset | Attribute | Detail | |-----------|--------| | Source | Open Africa Datasets | | Variable | Monthly Average Temperature (°C) | | Period | January 1991 – December 2016 | | Frequency | Monthly | | Observations | 312 | | Missing Values | None | --- ## Methodology ### 1. Exploratory Data Analysis * The series exhibits clear annual seasonality (12-month cycle) with a mild upward warming trend. * Additive seasonal decomposition confirms a stable seasonal amplitude of approximately ±1.5°C and largely random residuals, justifying the use of an additive modelling framework. ### 2. Stationarity & Preprocessing * The Augmented Dickey-Fuller (ADF) test returned a p-value of 0.3658, confirming the series is non-stationary. * First-order differencing (d = 1) was applied, after which the series became stationary and mean-reverting with no visible trend. ### 3. Model Development **SARIMA(3,1,1)(2,0,1)[12]** Selected via stepwise Auto ARIMA (AIC minimisation, AIC = 396.85), consistent with ACF/PACF analysis of the differenced series. | Component | Terms | |-----------|-------| | Non-seasonal | AR(3), I(1), MA(1) | | Seasonal | SAR(2), SMA(1), m = 12 | > Note: No seasonal differencing (D = 0) was applied. The seasonal structure is captured parametrically through the SAR and SMA terms, with the > SAR(L12) coefficient of 1.094 implying near-implicit seasonal differencing within the AR structure. **Holt-Winters Exponential Smoothing** * Included as an interpretable benchmark against the SARIMA model. * Fitted with additive trend and additive seasonality (period = 12), appropriate …